Add missing numba-cuda dependency to the conda recipe

[pentschev]

The numba-cuda package was made a dependency in #462, but the conda recipe was missed.

[jameslamb]

I'm unsure that this is the right fix, left some clarifying questions.

[jameslamb]

cudf is updating its pinning to >=0.18.0,<0.19.0 in rapidsai/cudf#19604, so this will need to be updated again when that merges (to ensure all of RAPIDS can solve together).

Would it be acceptible to do something like >=0.14.0,<0.19.0 here, so environments that end up with both cudf and ucxx aren't broken? And then tighten it to >=0.18.0,<0.19.0 once that cudf PR is merged? Or is that range too large?

Also, should it be a run: (required) dependency not run_constraints: (optional)? I see that for wheels we only have this as an optional dependency in the [test] extra:

ucxx/python/ucxx/pyproject.toml

Line 44 in d2d9468

"numba-cuda[cu12]>=0.14.0,<0.15.0a0",

But it seems looks like it's imported unconditionally:

ucxx/python/ucxx/examples/basic.py

Line 15 in d2d9468

import numba.cuda

[pentschev]

As a recap, we had to rollback numba-cuda version dependency twice before, immediately this is NOT acceptable IMO. I don’t know if you recall, here's how events unfolded:

1. First introduced dependency: d3aa4a3
2. Rolled back: 0f5f99f
3. And rolled back again: 41fa142

Since this is a critical issue, that we’ve been seeing segmentation faults in both RAFT and RMM, I think we first need to go the safe route and make sure everything is back to normal, then we can update the pin to a newer version.

[pentschev]

Moved to run section in a57320f .

[jameslamb]

This is looking good, thank you! Merge any time.

[pentschev]

Thanks @jameslamb and @KyleFromNVIDIA , triggering auto-merge.

[pentschev]

/merge

[bdice]

Trying to spell this consistently across RAPIDS to make automated updates easier.

Suggested change

	- &numba-cuda numba-cuda>=0.14.0,<0.15.0a0
	- &numba-cuda-dep numba-cuda>=0.14.0,<0.15.0a0

[pentschev]

@KyleFromNVIDIA @bdice you should agree here. See rapidsai/dask-cuda#1531 (comment) for the identical discussion in the opposite direction.

[KyleFromNVIDIA]

@bdice Most YAML anchors in dependencies.yaml use underscores and don't have -dep at the end. The existing numba-cuda ones are not consistent with everything else. We should fix the existing ones, not make even more inconsistent ones.

[bdice]

I applied my changes (minor renames / typos) since I just saw that /merge had already been triggered.

[pentschev]

Just for the sake of completeness here about #481 (comment), Kyle agreed we should go forward with the current naming and fix with rapids-reviser later if needed.